The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes text and instructions provided via CLI arguments and JSON files to generate speech. This content is passed to local TTS models without sanitization, creating a surface for potential indirect prompt injection attacks if the input data originates from untrusted sources.
Ingestion points: scripts/batch_dubbing.py loads speech segments from a JSON file path provided via the --input argument; scripts/run_tts.py accepts text and style instructions via --text and --instruct.
Boundary markers: There are no boundary markers or explicit instructions to the underlying models to ignore embedded directives in the processed text.
Capability inventory: The skill scripts have the capability to write audio files to user-specified directories (soundfile.write), perform directory cleanup (shutil.rmtree), and trigger network requests for model downloads via the Hugging Face Hub.
Sanitization: No input validation or filtering is performed on the input text before it is processed by the TTS engine.
[DYNAMIC_EXECUTION]: The skill performs runtime monkeypatching of the transformers library to modify its default behavior.
Evidence: scripts/batch_dubbing.py contains a configure_transformers function that overrides AutoTokenizer.from_pretrained with a local wrapper to inject the fix_mistral_regex parameter. While intended to address a compatibility issue, dynamic modification of library methods is an execution pattern that requires oversight.
[EXTERNAL_DOWNLOADS]: The skill is configured to download pre-trained models from the Hugging Face Hub during its first execution.
Evidence: The skill documentation and scripts reference models from the mlx-community organization on Hugging Face, such as Qwen3-TTS-12Hz-0.6B-CustomVoice-4bit. These downloads are triggered automatically by the mlx-audio library when the models are not present locally.

qwen3-tts-mlx